Forced Alignment Under Adverse Conditions

The problem of forced alignment is that of matching phonetic segments in an audio sample to its corresponding transcription, which is a vital part of indexing audio files. While various methods have been employed to accomplish this task, the results become less accurate under adverse alignment conditions caused by various disturbances in the audio as well as transcription errors. In fact, the alignment errors are usually left undiscovered until the aligned audio and transcript combination is later reviewed by human eyes and ears, thus defeating the purpose of an automated transcription and alignment process. This project seeks to develop a robust method to improve existing forced alignment techniques and increase their functionality. This will be accomplished by developing a technique to detect errors in alignment and produce correction algorithms to reduce the frequency of these errors. Various methods to find and fix errors in the alignment process will be examined. By combining these different techniques, a more accurate forced alignment package will be generated, which will be able to operate in adverse conditions found in both the transcript and audio.